Compare the best speech-to-text APIs for voice agents, call transcription, meeting intelligence, and multilingual ASR by latency, turn detection, accuracy, language coverage, and price.
Start with the use case: for Realtime voice agent with barge-in and turn-taking, pick Deepgram; for Voice agent where transcription quality beats cheapest streaming, pick AssemblyAI; for Want max accuracy or to self-host at scale, pick OpenAI Whisper; for Post-call analytics, meeting notes, or contact-center intelligence, pick AssemblyAI.
Flux is designed for conversational pipelines where knowing when a speaker is done matters as much as the transcript.
Pick DeepgramUniversal-3 Pro Streaming is positioned for higher-quality real-time transcription, with a cheaper streaming model available when cost matters more.
Pick AssemblyAIWhisper is the open-source accuracy gold standard, and self-hosting removes per-minute cost at high volume.
Pick OpenAI WhisperAssemblyAI's speech-understanding add-ons reduce the amount of downstream NLP you have to assemble yourself.
Pick AssemblyAIA side-by-side view of type, cloning, languages, commercial licensing, and benchmark notes — every price is dated with its official source.
| Tool | Type | Cloning | Free tier | Starting price | Languages | Commercial use | Latency / accuracy | Benchmark note | Checked |
|---|---|---|---|---|---|---|---|---|---|
| Deepgram | ASR | No | Yes | $0.0048/min | Nova supports 45+ languages; Flux has English and multilingual options | Commercial use under standard terms; self-hosted/on-prem available | Flux is built specifically for voice agents | Realtime voice-agent ASR | 2026-06-22 |
| AssemblyAI | ASR | No | Yes | $0.15/hr | Universal-2 supports 99 languages; Universal-3 Pro/Streaming currently covers English, Spanish, German, French, Italian, and Portuguese | Commercial use under standard API terms | Universal Streaming at $0.15/hour | Transcription intelligence | 2026-06-22 |
| OpenAI Whisper | ASR | No | Yes | Free (self-host) / $0.006/min API | 99+ languages incl. Chinese | MIT license — free for commercial use | Open repository and self-hosted control | Open or self-hosted ASR baseline | 2026-06-12 |
| Google Cloud Speech-to-Text | ASR | No | Yes | Usage-based | 125+ languages | Commercial use under Google Cloud terms | Streaming recognition returns results in real time over gRPC | GCP-native enterprise ASR | 2026-06-22 |
| ElevenLabs Scribe | ASR | No | Yes | Included in ElevenLabs plans | Multilingual real-time | Commercial use on paid ElevenLabs plans | Scribe v1/v2 speech-to-text at $0.22/hour | ElevenLabs voice-stack ASR | 2026-06-12 |
DeepgramVoice-agent defaultFlux is built for real-time agents with model-native turn detection, natural interruption handling, and about 260ms end-of-turn detection; Nova-3 covers high-accuracy general ASR.
Flux is built for real-time agents with model-native turn detection, natural interruption handling, and about 260ms end-of-turn detection; Nova-3 covers high-accuracy general ASR.
AssemblyAIQuality plus intelligenceUniversal-3 Pro Streaming targets higher-quality voice-agent transcription, while Universal-Streaming keeps real-time ASR at a lower $0.15/hr entry point; intelligence add-ons cover summaries, sentiment, and labels.
Universal-3 Pro Streaming targets higher-quality voice-agent transcription, while Universal-Streaming keeps real-time ASR at a lower $0.15/hr entry point; intelligence add-ons cover summaries, sentiment, and labels.
OpenAI WhisperAccuracy & open sourceStrong multilingual accuracy and self-hosting economics, but live voice agents need extra streaming, turn-taking, and diarization engineering around the model.
Strong multilingual accuracy and self-hosting economics, but live voice agents need extra streaming, turn-taking, and diarization engineering around the model.
Google Cloud Speech-to-TextBroad languagesEnterprise ASR on Google Cloud with streaming over gRPC and billing by successfully processed audio in one-second increments.
Enterprise ASR on Google Cloud with streaming over gRPC and billing by successfully processed audio in one-second increments.
ElevenLabs ScribeMultilingual realtimeAccurate multilingual transcription with real-time support, ideal if you already use ElevenLabs for TTS and want one vendor.
Accurate multilingual transcription with real-time support, ideal if you already use ElevenLabs for TTS and want one vendor.
The best speech-to-text tools and APIs include Deepgram, AssemblyAI, OpenAI Whisper, Google Cloud Speech-to-Text, and ElevenLabs Scribe. For real-time voice agents, start with Deepgram Flux when turn-taking, interruption handling, and first-response latency matter most. Pick AssemblyAI when you need transcript quality plus speech intelligence, Whisper when you want accuracy and self-hosting control, Google Speech-to-Text when your stack already lives on GCP, and ElevenLabs Scribe when you want ASR beside ElevenLabs TTS.
For voice agents, test turn detection, interruption handling, partial transcript speed, and first response latency before you compare generic WER numbers. For batch transcription, run a 30-minute sample set across clean calls, noisy calls, accents, and domain vocabulary before committing. Watch add-on pricing carefully: diarization, redaction, keyterm prompting, sentiment, and summaries can stack on top of the base rate. Separate voice-agent ASR from post-call analytics. The best real-time recognizer is not always the best meeting intelligence product.
Deepgram, AssemblyAI, OpenAI Whisper, Google Cloud Speech-to-Text, and ElevenLabs Scribe offer a usable free tier or free entry, so you can evaluate them without paying. Paid plans typically start around $0.0048/min.
Realtime voice agent with barge-in and turn-taking → Deepgram; Voice agent where transcription quality beats cheapest streaming → AssemblyAI; Want max accuracy or to self-host at scale → OpenAI Whisper; Post-call analytics, meeting notes, or contact-center intelligence → AssemblyAI; Enterprise team already standardized on Google Cloud → Google Cloud Speech-to-Text.